on discrimination discovery lu zhang using causal …sbp-brims.org/2016/program/f_250pm.pdfdirect...
Post on 28-May-2020
16 Views
Preview:
TRANSCRIPT
ON DISCRIMINATION DISCOVERY USING CAUSAL NETWORKS
Lu ZhangYongkai WuXintao WuUniversity of Arkansas
OUTLINE
Introduction
Related Work
Methods
Preliminary Results
Conclusions and Future Work
INTRODUCTION
Discrimination refers to unjustified distinctions of individuals based on their membership in a certain group.
Federal Laws and regulations disallow discrimination on several grounds: Gender, Age, Marital Status, Sexual Orientation, Race, Religion or Belief, Disability or
Illness …… These attributes are referred to as the protected attributes.
In 2014, US President Obama called for a 90-day review of big data collecting and analyzing practices. “Big data technologies can cause societal harms beyond damages to privacy, such as
discrimination against individuals and groups”.
INTRODUCTION
RELATED WORK
Several types of discrimination have proposed in the literature. Accordingly, different types discrimination discovery techniques have been developed. Group discrimination (association rules), individual discrimination (situation testing),
direct discrimination (conditional discrimination analysis), and indirect discrimination (rule inference), etc.
The down side: Current approaches are mainly based on correlation or association. In discrimination
discovery, it is critical to derive causal relationship, and not merely association relationship. We need a unifying framework for determining all types of discrimination rather than
using different types of data mining techniques.
METHODS
For a quantitative measurement, we use risk difference to measure discrimination. The difference in the proportion of positive decisions between the protected group
and the non-protected group. Formally, ∆P|s = P(e+|c+, s) − P(e+|c−, s). E: decision attribute (admission); e+: positive decision (admitted); e-: negative decision (rejected).
C: protected attributes (gender); c+: non-protected group (male); c-: protected group (female).
S: A set of attributes that represents a partition; s: a subpopulation under the partition.
Discriminatory effect presents if ∆P|s > τ, where τ is a threshold for discrimination depending on law (e.g., 5%).
The key to discrimination discovery: find meaningful partitions.
ILLUSTRATIVE EXAMPLES
In a university admission system: gender, major, test score, admission.
Overall admission rate:Female: 37%Male: 47%
Overall admission rate:Female: 43%Male: 43%
We use the causal network to find the meaningful partitions.
CAUSAL NETWORK (CAUSAL GRAPH, CAUSAL BAYESIAN NETWORK, ETC.)-BASED APPROACH
The causal network:
Causal effects are carried by the paths that trace arrows pointing from the cause to the effect.
In the causal network theory, a path can be blocked by conditioning on a set of nodes not containing the two end-nodes. Upon blocked, the effect originally transmitted through the path is suppressed in each
subpopulation under the partition defined by the set of nodes. Use path blocking to identify the causal effects that we want to measure for discrimination
discovery.
The criterion of blocking a path is given in the d-separation criterion. E.g., the path C → Z → E is blocked when conditioning on Z.
Protected attribute
Decisionattribute
DISCRIMINATION CATEGORIZATION
Different criteria are needed for different types of discrimination.
We categorize discrimination into:
system-level (e.g., against the university)
group-level (e.g., against a
department)
individual-level
direct discrimination
two criteria
indirect discrimination
SYSTEM-LEVEL DIRECT DISCRIMINATIONDirect: the direct causal effect carried by C → E. The direct causal effect can be identified under the partition determined by a certain
set of attributes, which we refer to as the block set.1. All paths other than C → E are blocked.
2. Does not contain E’s decedents.
For each block set B (defines a partition) and its value assignment b (defines a subpopulation), ∆P|b = P(e+|c+, b) − P(e+|c−, b) can be used to measure the direct causal effect.
System-level: the discriminatory effect should be measured as a persistent negative effect under all meaningful partitions. Discriminatory effect needs to measured for each value assignment b of each block
set B.
SYSTEM-LEVEL DIRECT DISCRIMINATION
Rigorous discrimination criterion: Discrimination is considered to present if inequality ∆P|b > τ holds for each value assignment b of each block set B. In real situations, ∆P|b may vary due to randomness and small sample size.
Relax (τ,α)-discrimination criterion: Discrimination is considered to present if Pr( ∆P|B > τ ) > α, where α is a user-defined parameter. Can be estimated using the Cantelli’s inequality: Pr( ∆P|B > τ ) > α if
μα
1 α ∙ σ τ
SYSTEM-LEVEL DIRECT DISCRIMINATION
The previous two criteria require to examine all block sets. A brute force algorithm of finding all block sets has an exponential complexity.
It is proved that, it is sufficient to examine Q instead of examining each block set, where Q is the set of E’s all parents except C.
Rigorous discrimination criterion: Discrimination is considered to present if inequality ∆P|q > τ holds for each value assignment q of set Q.
Relax (τ,α)-discrimination criterion: Discrimination is considered to present if
μα
1 α ∙ σ τ
SYSTEM-LEVEL INDIRECT DISCRIMINATIONConcerns about the indirect causal effects that are transmitted through intermediate attributes along the causal paths from C to E other than the direct arc C → E. Not all of them.
We refer to the attributes on the causal paths whose usage cannot be legally justified as the redlining attributes.
We propose to measure the indirect causal effect through the paths that each contains at least one redlining attribute.
Similarly, system-level indirect discrimination can be claimed if persistent negative effects are measured.
GROUP-LEVEL DIRECT DISCRIMINATION
Occurs in a particular subsystem other than across the whole system. The group G can be specified by analysts to denote a subsystem, determined by a
subset of profiling attributes. E.g., when we determine whether there exists group-level discrimination in a
particular major (e.g., CS) in university admission, G contains all applicants in CS.
The determination of block set B needs to be adjusted based on the given group G to form a partition within the given group. E.g., when focusing on group-level discrimination in CS major, B may contain test
scores. For each test score b, group-level discrimination can be claimed after we examine ∆P|b across all test scores among CS applicants.
INDIVIDUAL-LEVEL DIRECT DISCRIMINATIONTo identify discrimination for a specific individual, i.e., an entire record in the dataset.
Find two neighborhood groups that contain similar individuals from the protected group and the non-protected group. The individual is considered as discriminated if significant difference is observed between the decisions from the two groups.
Use the causal network to find similar individuals.
The causal structure and the causal effect of each attribute on the decision attribute are used for the calculation of inter-individual distances.
PRELIMINARY RESULTS
Apply the rigorous discrimination criterion and the relax (τ,α)‐discrimination criterion to discovery system-level direct discrimination. Datasets: Adult dataset and Dutch Census 2001 The causal network is built using the PC algorithm on TETRAD. τ is set as 0.05. α is set as 60%.
PRELIMINARY RESULTS (ADULT)
Protected and decision attributes
Parents of decision attribute
Rigorous: the values of ∆P|q range from -0.614 to 0.524Relaxed: Pr( ∆P|Q > 0.05 ) > 0.09% according to the Cantelli’sinequality
PRELIMINARY RESULTS (DUTCH)
Protected and decision attributes
Parents of decision attribute
Rigorous: the values of ∆P|q range from 0.062 to 0.435Relaxed: Pr( ∆P|Q > 0.05 ) > 65.48% according to the Cantelli’sinequality
CONCLUSIONS AND FUTURE WORK
Studied the problem of discrimination discovery.
Categorize different discrimination types based on discrimination scale and discrimination manner.
Based on the causal network, established two criteria for system-level direct discrimination: a rigorous criterion and relaxed criterion.
In the future work: extend the results to other types of discrimination, and study the problem of discrimination prevention, which aim to remove discrimination by modifying the data before conducting predictive analysis.
THANK YOU!
Acknowledgment This work was supported in part by U.S. National Institute of Health
(1R01GM103309).
Contact Lu Zhang, Email: lz006@uark.edu
top related